Compete, Don't Erase: Contradictory Fine-Tuning Masks Rather Than Deletes Knowledge

Does fine-tuning a language model on a fact that contradicts what it already knows delete the old value, or merely let a new value out-compete it? Reading the model's own log-probabilities of the old and the new value across training answers it directly: across three models in two families (Qwen2.5-1.5B/3B and Llama-3.2-3B, the latter on two independent fact draws), contradictory fine-tuning leaves the old value within ~1-1.5 nats of its pre-contradiction reference — zero of the contradicted facts are erased — while the new value wins by rising, not by the old one falling. The same readout, applied to a localized weight edit (constrained FT-L on a different model, GPT-2 XL), instead reads erasure (log-probability of the old value collapses ~20 nats, no return under paraphrase), so the diagnostic discriminates: it reports an erasure when a real erasure occurs. The new value is up-weighted in proportion to the surprise of the correction (the error-proportional cross-entropy signature made directly visible); a stronger prior resists correction more (a moderate, per-fact reactance effect); and an across-layer tuned-lens reading localizes the change as a shallow late patch on an intact old-value trunk — the early and middle layers still compute the old value, and the new value is written almost entirely into the final layers. On the masked model the old value behaves like masked, not deleted, content: under a matched budget it is relearned several times faster than it was first installed, and a context shift recovers part of its margin (the rapid-reacquisition and renewal signatures documented for extinction in behavioural neuroscience). A pre-registered per-fact prediction about recovery reversed and is reported as a disconfirmation. The contribution is a lens, a validated mask-versus-erase diagnostic (with a continuous, threshold-insensitive erase index), and a cross-substrate unification — not a new unlearning method and no state-of-the-art forgetting number. The readout is the direct mask-versus-erase metric the unlearning literature has explicitly said it lacks. The competing-routes substrate is developed in the series' foundational papers, and the confident-wrong-at-the-near-tie reading is shared with the measurement-model companion. Data and code. The synthetic-fact generators, training and evaluation scripts, and per-checkpoint logit trajectories are released with the publication in the companion repository: https://github.com/tplund/friction-theory-p2f-compete-not-erase. v2 (August 2026) — prior-art and attribution revision. Three titled precedents in the machine-unlearning literature are now cited and engaged, including one published three weeks before v1 that addressed the same evaluation gap; the novelty claim is narrowed accordingly, from the question itself to the direct output-distribution-level measurement of it. The distinction the paper turns on is given the name it has carried in human memory research since 1992: storage strength versus retrieval strength, and availability versus accessibility before that. The paper's own quantity is stated as an operationalisation of that construct rather than as an identification with it, and the asymmetry is made explicit — storage strength is not directly observed here, and its closest analogue is the savings-under-relearning measure the theory itself names. A prediction that failed pre-registration is still reported as failed, but the direction it went is now situated in the human literature that finds the same direction, rather than left as a bare defeat. A mis-stated article title is corrected, a reference cited in the body but absent from the list is added, and a set of distances quoted against the paper's own Table 1 is corrected at every occurrence. Companion citations updated to their published records. Editorial pass. Earlier versions remain in the version history.